Lightweight Neural Network for Sleep Posture Classification Using Pressure Sensing Mat at Various Sensor Densities.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Wearable Technology

Background:

• Pressure-sensing mats are used for sleep posture recognition, but face challenges balancing computational complexity and performance.
• Existing systems struggle with real-time monitoring due to high computational demands and potential time latency.

Purpose of the Study:

• To propose a lightweight neural network, ConcatNet, for efficient and accurate real-time sleep posture recognition.
• To investigate the impact of sensor density on recognition performance using ConcatNet models of varying scales.

Main Methods:

• Developed ConcatNet, incorporating inception modules for multi-receptive field feature extraction and multi-layer feature fusion.
• Utilized depthwise convolution to enhance model efficiency.
• Evaluated ConcatNet models (ConcatNet-S, ConcatNet-M, ConcatNet-L) with different sensor densities.

Main Results:

• ConcatNet-M, with medium sensor density (16x16), achieved the highest performance.
• Achieved 95.56% accuracy in short-term cross-validation and 94.68% in overnight testing.
• Demonstrated minimal model size (7.91KB), FLOPs (56.47K), and inference time (0.38ms).

Conclusions:

• ConcatNet offers a highly efficient solution for real-time sleep posture recognition.
• The model's low computational requirements make it suitable for deployment on mobile devices.
• This research addresses the trade-off between computational complexity and performance in pressure-sensing mat systems.